P. Marchetta , P. Mérindol , B. Donnet A. Pescapé , J.-J....
Transcript of P. Marchetta , P. Mérindol , B. Donnet A. Pescapé , J.-J....
P. MarchettaI, P. MérindolII, B. DonnetIII,
A. Pescapé I, J.-J. Pansiot II
I. University of Napoli “Federico II”, Italy
II. University of Strasbourg, France
III. University of Liege, Belgium
Topology Discovery and IGMP probing
IGMP filtering
Quantifying the IGMP filtering impact
Reassembling strategy
Sprint Network. Router-level Map
MERLIN project. 2011
The Internet topology as a dynamic graph of
IP Interfaceso Traceroute
Routerso Traceroute & Alias Resolution, IGMP
Point of Presence
Autonomous Systems (AS)o IP-2-AS and Router-2-AS mapping
Goalso IP network models & simulations
o Ground truth input for topology generation
Sprint Network. Router-level Map
MERLIN project. 2011
The Internet topology as a dynamic graph of
IP Interfaceso Traceroute
Routerso Traceroute & Alias Resolution, IGMP
Point of Presence
Autonomous Systems (AS)o IP-2-AS and Router-2-AS mapping
Goalso IP network models & simulations
o Ground truth input for topology generation
IGMP ASK-FOR-NEIGHBORS probes to elicit IGMP NEIGHBORS-
REPLY messages.
The targeted router provides information about its multicast
enabled interfaces and the links involving those interfaces.
The ME.R.L.IN project
Natively discovering network at Router level
Recursively exchange of ASK-FOR-NEIGHBORS and NEIGHBORS-REPLY messages
Advantageso Highly accurate: no inference.
o Alias Resolution: no need to gather IP interfaces.
o Network friendly: 1 probe injected per router.
o Forwarding independent: backup links reported.
o Layer-2 infrastructure partially inferred.
Natively discovering network at Router level
Recursively exchange of ASK-FOR-NEIGHBORS and NEIGHBORS-REPLY messages
Drawbackso Multicast scope
o IGMP local and in-transit filtering
Natively discovering network at Router level
Recursively exchange of ASK-FOR-NEIGHBORS and NEIGHBORS-REPLY messages
Drawbackso Multicast scope
o IGMP local and in-transit filtering
Local filtering
o The targeted router silently discards the packet probe.
In-transit filtering◦ The IGMP probe or its reply are dropped along the path.
Even a single not responding routermay induce a great fragmentation!
MERLIN-based experimental campaign toward
Sprint, Level3, Global Crossing.
MERLIN monitor
MERLIN server
1. Few large connected components2. Most isolated replying routers
Analyzing the hybrid graph (IGMP + Traceroute)
Nodes
IGMP replying routers
IPsIGMP
ICMP (Traceroute)
Links
IGMP – IGMP
ICMP – ICMP
IGMP – ICMP
Graph reduction1. Collapse connected IGMP routers in a single node (IGMP
connected component – c.c.)
2. Set the weight of each link to 1
3. Identify IP nodes with a degree>=3
Graph reduction1. Collapse connected IGMP routers in a single node (IGMP
connected component – c.c.)
2. Set the weight of each link to 1
3. Identify IP nodes with a degree>=3
Graph reduction1. Collapse connected IGMP routers in a single node (IGMP
connected component – c.c.)
2. Set the weight of each link to 1
3. Identify IP nodes with a degree>=3
4. Remove IP nodes with degree of 2
5. Keep track of the original distances with
new weights
Graph reduction1. Collapse connected IGMP routers in a single node (IGMP
connected component – c.c.)
2. Set the weight of each link to 1
3. Identify IP nodes with a degree>=3
4. Remove IP nodes with degree of 2
5. Keep track of the original distances with
new weights
Graph reduction1. Collapse connected IGMP routers in a single node (IGMP
connected component – c.c.)
2. Set the weight of each link to 1
3. Identify IP nodes with a degree>=3
4. Remove IP nodes with degree of 2
5. Keep track of the original distances with
new weights
6. Compute the shortest path for
each pair of IGMP c.c. (Dijkstra)
7. Compute the minimal weighted tree
(Kruskal) ({A,B,C} {A-B, B-C})
Graph reduction1. Collapse connected IGMP routers in a single node (IGMP
connected component – c.c.)
2. Set the weight of each link to 1
3. Identify IP nodes with a degree>=3
4. Remove IP nodes with degree of 2
5. Keep track of the original distances with
new weights
6. Compute the shortest path for
each pair of IGMP c.c. (Dijkstra)
7. Compute the minimal weighted tree
(Kruskal) ({A,B,C} {A-B, B-C})
How far are the IGMP components from each other?
Before applying Kruskal After applying Kruskal
All the links in the minimal weightedtree have a weight of 2!
3. There exists at least one path between most pairs of IGMP c.c.
4. Each IGMP component is not locatedfarther than 2 hops from
its closest component
How to obtain a full connected router-level topology of the network?
Alias resolution technique to transform the hybrid graph in a full router level topology.
However ◦ Alias resolution techniques are intrusive, time costing, and error-prone;
◦ IGMP pure topology is highly accurate.
How to preserve the IGMP accuracy providing a full connected router level graph?
Apply Alias Resolution (Ally) to transform the hybrid graph in a
router level topology.
IGMP router
ICMP IP
IGMP IP
Apply Alias Resolution (Ally) to transform the hybrid graph in a
router level topology.
IGMP router
ICMP IP
IGMP IP
Applying alias resolution on the entireIP level portion of the topology is extremely
time-costing and error-prone!
IGMP router
ICMP IP
IGMP IP
Key Idea: consider only the IPs located close to the routers in the
current router level topology.
Apply alias resolution on the IP level neighborhoodof current router level topology.
IGMP router
ICMP IP
IGMP IP
Apply alias resolution on the IP level neighborhoodof current router level topology.
Key Idea: consider only the IPs located close to the routers in the
current router level topology.
IGMP router
ICMP IP
IGMP IP
Aliased routerGuadually expand each connected component
iteration by iterationuntil the router level graph is full connected.
Key Idea: consider only the IPs located close to the routers in the
current router level topology.
Disjoint Componentsevolution
Links EvolutionNodes Evolution
5. Strong components reduction
6. Largest components stronglygrow (Sprint 393 nodes)
7. Strong reduction of the aliasingspace exploration ( -95%)
AS #Components
Sprint - 91%
Global Crox. - 38%
Level3 - 62%
After 2 iterations
IGMP filtering causes the collected topologies to be disconnected.
We proposed
◦ A hybrid graph reduction method to investigate and characterize the
phenomenon;
◦ An efficient reassembling strategy able to strongly reduce the number of
components.
For the first time in literature, we jointly exploited IGMP probing,
Traceroute and Alias resolution in Topology Discovery.
Our topologies are freely available at
http://svnet.u-strasbg.fr/merlin
Multiple vantage points allow to
deal with the in-transit filtering
Local filtering is still challenging.
Sprint Network.The final IGMP topologies consist of
several disjoint components!
#components component size
IGMP filtering is getting worseand worse!
2006 2007 2008 2009